MOD · Model

AI agents and LLM automation

I design AI agents and model orchestration that use company tools and data, with least-privilege permissions, measurable evaluations and human oversight.

Let's talk

Who it is for

Companies that want to put an agent in production (support, documents, internal automation) without first building a dedicated department.

Where

Pavia and Lombardy on site by appointment, across Italy and Europe remotely.

How it starts

A call, then a written proposal with scope, timing and cost.

Problems I solve

• Prototypes that never reach production: no evaluations, tracing or cost limits.
• Security risk: prompt injection and agents with too many permissions (OWASP LLM01 and LLM06).
• Integrating with existing systems: tools exposed through MCP, flows with human approval.

What you get

• Agent architecture (orchestrator, tools, memory) and model selection
• Automated evaluation suite and tracing with OpenTelemetry
• Permission model and a prompt injection mitigation plan

In practice

I design AI agents that work with a company's tools and data: document search (RAG), tool use (function calling, MCP), flows with human approval. Every agent gets minimal permissions, capped costs and a measurable evaluation before it goes to production.

• From use case to prototype in a few weeks
• Automated evaluations, logs and tracing
• Least-privilege permissions, prompt injection defences
• Running on your own or managed infrastructure

Technologies

• MCP
• Claude
• OpenAI
• RAG
• OpenTelemetry GenAI
• OWASP LLM Top 10
• Kubernetes

How I work

1. Listening A call to understand the context, the problem and the constraints. No commitment.
2. Proposal Scope, timing and cost written clearly, before anything starts.
3. Delivery Small shippable steps, regular updates and code in your own repository.
4. Handover Documentation and training: the work ends when your team can run it alone.